From the number one player, to the game breaker
Source| AI Blue Media Exchange
ID:lanmeih001
Author|Yiqi
Editor|Wei Xiao
In March 2023, NVIDIA CEO Jensen Huang wore a black jacket and stood on the podium of the company's spring GTC conference, saying the phrase "ChatGPT is equivalent to the iPhone in the AI industry." ”
Lao Huang believes in ChatGPT, and the outside world believes in Lao Huang, and as soon as this word came out, it was widely circulated.
From the tone of speech, Huang Jenxun's words are not intended to provide the industry with an open point of view, he is really "informing" the world that the AI era belonging to ChatGPT has come.
"I think that's a foregone conclusion."
Of course, Lao Huang has the confidence to say so.
Someone once talked to Huang Jenxun about NVIDIA's AI layout, and Huang said that NVIDIA discovered that AI can "change everything" ten years ago, and then the company has always regarded AI technology as the future from the bottom to the top. "Every chip we make, it's focused on artificial intelligence."
Lao Huang also has every reason to say so.
The communication speed between CPU-memory-GPU has been the bottleneck of AI computing for a long time. Among all the options available in the current market, the data transfer speed of NVIDIA products is far ahead of all competitors, including AMD and Intel. Vivek Arya, a semiconductor analyst at Bank of America Securities, said, "It's very easy to use NVIDIA's products and add computing power, which is basically equivalent to hard currency in Silicon Valley." ”
Nvidia reached the publisher of hard currency, and Huang Jenxun also went from being the "number one player" of AI to the one who set the rules of the game in the ChatGPT era.
Admission ticket to the ChatGPT era
On the NVIDIA website, you can find this sentence: "In 1999, NVIDIA invented the GPU. ”
It is not so much that AI chose NVIDIA, but that AI chose the GPU (graphics processing unit), and NVIDIA is the creator of GPUs and the number one player in the GPU industry.
In the nineties of the 20th century, the gaming industry suffered from insufficient CPU processing power for geometry to further unleash performance.
NVIDIA was the first to find an understanding.
In 1997, NVIDIA launched the world's first 128-bit 3D processor, the RIVA 128, with shipments exceeding one million units within four months of its release, stepping on the "big step for mankind" on the moon soil of graphics chips.
Two years later, in 1999, the world's first GPU, the GeForce 256, was released, also made by NVIDIA.
Unlike CPUs (central processing units), which have always performed complex calculations in chronological order, NVIDIA GPUs work with hundreds of cores to process the pixels of each graphics one-to-one. All cores work at the same time, and all pixels are processed at the same time.
For simple processes and large amounts of data, the emergence of GPUs has brought exponential efficiency improvements. Since then, NVIDIA graphics chips have been equivalent to GPUs, and they are the best CPUs you can find on the planet.
Nvidia is the one who paves the way for AI. Twenty years later, AI will need exactly this structure to support a large number of computing cores—the most common convolution operation in the AI field, which is essentially a series of combinations of addition, subtraction, multiplication and division.
AI deep learning relies on computations that are dense and can be executed in large numbers in parallel, and GPUs excel at exactly this kind of computing. Since then, NVIDIA has naturally become the first choice for most AI companies.
According to relevant data, 80.6% of the world's cloud computing and data centers in 2020 were powered by NVIDIA GPU; In 2021, China's domestic GPU servers accounted for more than 88.4% of the domestic server market, and NVIDIA GPUs accounted for more than eighty percent of them.
In the third quarter of 2022, Jon Peddie Research released a report that Q3 discrete GPU shipments that year were 14 million units, of which 88% of the market share belonged to NVIDIA.
In the era of AI with tickets, NVIDIA is printed on every ticket.
In August 2016, NVIDIA donated the world's first AI supercomputer to OpenAI, which was less than a year old at the time, and wrote on the machine: "To Elon and the OpenAI team, for the future of computing and mankind, I donate the world's first DGX-1." ”
After the release of ChatGPT-3 in November 2022, Microsoft, the gold behind OpenAI, said that the company's Azure cloud service built an AI computing cluster with more than 10,000 NVIDIA A100 GPU chips for ChatGPT to provide computing power support for the development of ChatGPT.
But the landlord's family did not have much surplus food.
According to sources, Microsoft began to implement GPU resource quota supply internally at the end of 2022, but the approval time has become longer and longer since January this year, and some applications need to wait for days or even weeks to be approved. The tens of thousands of GPUs that Microsoft previously ordered from NVIDIA have also arrived indefinitely due to the shortage of market conditions.
The next round of landlords, Huang Jenxun, robbed.
Winner Lao Huang
Shakespeare wrote in Othello that jealousy is a green-eyed demon, and whoever falls into his trap will be played with by him.
Envy comes from the Latin word "invidia," the NVIDIA logo, which is a green eye.
This is no coincidence.
Jensen Huang and his Nvidia have indeed lived as jealous winners – Nvidia's AI supercomputer, DGX, is the engine behind the big model of language. After Jensen Huang handed over the world's first DGX to OpenAI, more than half of the world's top 100 companies also installed supercomputers from NVIDIA.
According to New Street Research, NVIDIA GPUs already account for 95% of the graphics processor market that can be used for machine learning.
To support Twitter's own AIGC project, Musk recently purchased about 10,000 GPUs. In March 2023, TrendForce, a US market research institution, released a report that GPT-3.5 large models that process 180 billion parameters require 20,000 GPU chips. In the future, the number of GPU chips required for the commercialization of GPT large models will exceed 30,000.
There is a saying in the cloud computing industry: 10,000 NVIDIA A100 chips are the threshold for computing power to do a good job in AI large models.
For some reasons, domestic companies can hardly get new A100 and H100 chips, and can only purchase A800 and H800 chips as replacements.
Relevant institutions estimate that there are only about 30,000 A100 chips in China, and the A100 replacement version of A800 launched by NVIDIA is also "difficult to find". It is reported that the price of an A100 GPU by some merchants has risen from about 60,000 US dollars to 90,000 or even 100,000 US dollars. The relatively poor A800 chip has also cost more than 80,000 US dollars.
There was a report in the market that reported that Baidu urgently placed an order for 3,000 A800 servers containing 8 chips (equivalent to 24,000 A800 chips) at the beginning of this year, and it is expected that there will be a total of 50,000 A800 and H800 in demand throughout the year. According to sources, in the sprint stage of the development process of Baidu's large model "Wen Xin Yiyan", almost all of the group's A100 chips were occupied.
Alibaba Cloud is also expected to have about 10,000 chips this year, of which 6,000 are H800. In addition, Alibaba Cloud will also purchase self-developed chips from its semiconductor company Pingtou Ge, which will be about 3,000 per year. And these orders containing tens of thousands of A800 and H800 chips, the single amount will not be less than one billion dollars.
Snapping up also happens in the car market.
He Xiaopeng once relayed an article titled "NVIDIA A100 restriction will hit China's autonomous driving hard" in the circle of friends, the article said that "the shortage of chips will bring challenges to the autonomous driving industry", and He Xiaopeng also admitted in the accompanying text that this is bad news.
"But the good news is that we have just bought back the demand for the next few years in advance", why do you think Xiaopeng is to keep a step, in fact, he also wants to put friendly businessmen in the army, "We will overcome all difficulties, will obviously surpass the next generation of friendly companies' 'fully autopilot' next-generation automatic driving assistance system, and next year, the country will really land." ”
He Xiaopeng had a fist expression, as if he was expressing determination, and as if he wanted to hit someone.
From the original graphics processors to the graphics myth, from the Bitcoin mining boom to today's ChatGPT, you'd be hard-pressed to find a second person like Jensen Huang who has been a "consistent winner." In April 2023, NVIDIA's total market value has reached $669.469 billion, surpassing TSMC to rank first among semiconductor companies, equivalent to five former chip overlords Intel. The company's share price rose by 90.05% in a single quarter (January to March 2023), and Huang's personal worth reached 24.456 billion US dollars.
The agency estimates that the AI industry will be an "$800 billion market opportunity" in the next few years.
With the release of the H100 chip, which is six times the performance of the A100 chip, Huang announced in March 2023 that NVIDIA will launch a new cloud leasing service - leasing supercomputers to the B side for the development of artificial intelligence technologies such as ChatGPT.
It seems to be open source, but it is actually expensive - the price of leasing this 8 A100 or H100 flagship chips is 37,000 US dollars / month, which is about 254,000 yuan.
In the ChatGPT era, Lao Huang, who can't see the pursuer, is still racing all the way...